Gray-box modeling, integrating physical principles with empirical data, offers a distinct advantage over black-box methods in equipment emulator development, particularly in terms of extrapolation beyond the training domain. Despite this, gray-box models often grapple with challenges such as nonlinearities, unmodeled dynamics, and susceptibility to local minima in the optimization landscape, which can undermine their performance. This paper introduces a novel approach to mitigate these issues by injecting controlled noise into the training dataset. This technique not only enriches the dataset but also imparts enhanced resilience to the model, enabling it to better handle inherent uncertainties. We demonstrate this methodology using a dynamic model of a water-to-water heat exchanger, evaluating its performance with live data streaming from real equipment. The results are compelling: noise injection led to a significant reduction in modeling error, with the root mean square error decreasing from 0.68 to 0.27 °C. This represents a 60% improvement in model accuracy on the training set, and notable enhancements of 50% and 45% on the test and validation sets, respectively. These findings not only underscore the efficacy of noise injection in gray-box modeling but also highlight its potential as an effective solution for enhancing the fidelity and reliability of equipment emulators in various applications.

This paper explores the application of autoencoder algorithms in Automated Fault Detection (AFD) for Heating, Ventilation, and Air Conditioning (HVAC) systems, specifically focusing on Fan Coil Units (FCUs). The study begins by reviewing the current state of Fault Detection and Diagnostics (FDD), emphasizing the limitations and the potential of unsupervised learning techniques like autoencoders and transfer learning to fill these gaps. Using data from a full-scale building case study featuring five Fan Coil Units (FCUs), the research develops and evaluates autoencoder-based AFD models that models effectively compress multivariate inputs into a reduced latent space, enabling accurate and efficient fault detection. The paper makes two novel contributions: (1) It introduces a methodology to distinguish between equipment-level and system-level faults; and (2) It demonstrates the generalizability of the approach across different types of FCUs through cross-testing and transfer learning. The results indicate that autoencoders outperform other dimensionality reduction algorithms and separate predictors in fault detection accuracy and efficiency. The paper concludes by discussing the implications of these findings for future research and practical applications in building management.